1 //===-- X86JITInfo.cpp - Implement the JIT interfaces for the X86 target --===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the JIT interfaces for the X86 target.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "jit"
15 #include "X86JITInfo.h"
16 #include "X86Relocations.h"
17 #include "X86Subtarget.h"
18 #include "llvm/CodeGen/MachineCodeEmitter.h"
19 #include "llvm/Config/alloca.h"
24 extern "C" void *_AddressOfReturnAddress(void);
25 #pragma intrinsic(_AddressOfReturnAddress)
28 void X86JITInfo::replaceMachineCodeForFunction(void *Old, void *New) {
29 unsigned char *OldByte = (unsigned char *)Old;
30 *OldByte++ = 0xE9; // Emit JMP opcode.
31 unsigned *OldWord = (unsigned *)OldByte;
32 unsigned NewAddr = (intptr_t)New;
33 unsigned OldAddr = (intptr_t)OldWord;
34 *OldWord = NewAddr - OldAddr - 4; // Emit PC-relative addr of New code.
38 /// JITCompilerFunction - This contains the address of the JIT function used to
39 /// compile a function lazily.
40 static TargetJITInfo::JITCompilerFn JITCompilerFunction;
42 // Get the ASMPREFIX for the current host. This is often '_'.
43 #ifndef __USER_LABEL_PREFIX__
44 #define __USER_LABEL_PREFIX__
46 #define GETASMPREFIX2(X) #X
47 #define GETASMPREFIX(X) GETASMPREFIX2(X)
48 #define ASMPREFIX GETASMPREFIX(__USER_LABEL_PREFIX__)
50 // Provide a convenient way for disabling usage of CFI directives.
51 // This is needed for old/broken assemblers (for example, gas on
52 // Darwin is pretty old and doesn't support these directives)
53 #if defined(__APPLE__)
59 // Provide a wrapper for X86CompilationCallback2 that saves non-traditional
60 // callee saved registers, for the fastcc calling convention.
62 #if defined(__x86_64__)
63 // No need to save EAX/EDX for X86-64.
64 void X86CompilationCallback(void);
68 ".globl " ASMPREFIX "X86CompilationCallback\n"
69 ASMPREFIX "X86CompilationCallback:\n"
70 CFI(".cfi_startproc\n")
73 CFI(".cfi_def_cfa_offset 16\n")
74 CFI(".cfi_offset %rbp, -16\n")
77 CFI(".cfi_def_cfa_register %rbp\n")
78 // Save all int arg registers
80 CFI(".cfi_rel_offset %rdi, 0\n")
82 CFI(".cfi_rel_offset %rsi, 8\n")
84 CFI(".cfi_rel_offset %rdx, 16\n")
86 CFI(".cfi_rel_offset %rcx, 24\n")
88 CFI(".cfi_rel_offset %r8, 32\n")
90 CFI(".cfi_rel_offset %r9, 40\n")
91 // Align stack on 16-byte boundary. ESP might not be properly aligned
92 // (8 byte) if this is called from an indirect stub.
94 // Save all XMM arg registers
96 "movaps %xmm0, (%rsp)\n"
97 "movaps %xmm1, 16(%rsp)\n"
98 "movaps %xmm2, 32(%rsp)\n"
99 "movaps %xmm3, 48(%rsp)\n"
100 "movaps %xmm4, 64(%rsp)\n"
101 "movaps %xmm5, 80(%rsp)\n"
102 "movaps %xmm6, 96(%rsp)\n"
103 "movaps %xmm7, 112(%rsp)\n"
105 "movq %rbp, %rdi\n" // Pass prev frame and return address
106 "movq 8(%rbp), %rsi\n"
107 "call " ASMPREFIX "X86CompilationCallback2\n"
108 // Restore all XMM arg registers
109 "movaps 112(%rsp), %xmm7\n"
110 "movaps 96(%rsp), %xmm6\n"
111 "movaps 80(%rsp), %xmm5\n"
112 "movaps 64(%rsp), %xmm4\n"
113 "movaps 48(%rsp), %xmm3\n"
114 "movaps 32(%rsp), %xmm2\n"
115 "movaps 16(%rsp), %xmm1\n"
116 "movaps (%rsp), %xmm0\n"
119 CFI(".cfi_def_cfa_register %rsp\n")
120 // Restore all int arg registers
122 CFI(".cfi_adjust_cfa_offset 48\n")
124 CFI(".cfi_adjust_cfa_offset -8\n")
125 CFI(".cfi_restore %r9\n")
127 CFI(".cfi_adjust_cfa_offset -8\n")
128 CFI(".cfi_restore %r8\n")
130 CFI(".cfi_adjust_cfa_offset -8\n")
131 CFI(".cfi_restore %rcx\n")
133 CFI(".cfi_adjust_cfa_offset -8\n")
134 CFI(".cfi_restore %rdx\n")
136 CFI(".cfi_adjust_cfa_offset -8\n")
137 CFI(".cfi_restore %rsi\n")
139 CFI(".cfi_adjust_cfa_offset -8\n")
140 CFI(".cfi_restore %rdi\n")
143 CFI(".cfi_adjust_cfa_offset -8\n")
144 CFI(".cfi_restore %rbp\n")
146 CFI(".cfi_endproc\n")
148 #elif defined(__i386__) || defined(i386) || defined(_M_IX86)
150 void X86CompilationCallback(void);
154 ".globl " ASMPREFIX "X86CompilationCallback\n"
155 ASMPREFIX "X86CompilationCallback:\n"
156 CFI(".cfi_startproc\n")
158 CFI(".cfi_def_cfa_offset 8\n")
159 CFI(".cfi_offset %ebp, -8\n")
160 "movl %esp, %ebp\n" // Standard prologue
161 CFI(".cfi_def_cfa_register %ebp\n")
163 CFI(".cfi_rel_offset %eax, 0\n")
164 "pushl %edx\n" // Save EAX/EDX/ECX
165 CFI(".cfi_rel_offset %edx, 4\n")
167 CFI(".cfi_rel_offset %ecx, 8\n")
168 #if defined(__APPLE__)
169 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
172 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
173 "movl %eax, 4(%esp)\n"
174 "movl %ebp, (%esp)\n"
175 "call " ASMPREFIX "X86CompilationCallback2\n"
176 "movl %ebp, %esp\n" // Restore ESP
177 CFI(".cfi_def_cfa_register %esp\n")
179 CFI(".cfi_adjust_cfa_offset 12\n")
181 CFI(".cfi_adjust_cfa_offset -4\n")
182 CFI(".cfi_restore %ecx\n")
184 CFI(".cfi_adjust_cfa_offset -4\n")
185 CFI(".cfi_restore %edx\n")
187 CFI(".cfi_adjust_cfa_offset -4\n")
188 CFI(".cfi_restore %eax\n")
190 CFI(".cfi_adjust_cfa_offset -4\n")
191 CFI(".cfi_restore %ebp\n")
193 CFI(".cfi_endproc\n")
196 // Same as X86CompilationCallback but also saves XMM argument registers.
197 void X86CompilationCallback_SSE(void);
201 ".globl " ASMPREFIX "X86CompilationCallback_SSE\n"
202 ASMPREFIX "X86CompilationCallback_SSE:\n"
203 CFI(".cfi_startproc\n")
205 CFI(".cfi_def_cfa_offset 8\n")
206 CFI(".cfi_offset %ebp, -8\n")
207 "movl %esp, %ebp\n" // Standard prologue
208 CFI(".cfi_def_cfa_register %ebp\n")
210 CFI(".cfi_rel_offset %eax, 0\n")
211 "pushl %edx\n" // Save EAX/EDX/ECX
212 CFI(".cfi_rel_offset %edx, 4\n")
214 CFI(".cfi_rel_offset %ecx, 8\n")
215 "andl $-16, %esp\n" // Align ESP on 16-byte boundary
216 // Save all XMM arg registers
218 // FIXME: provide frame move information for xmm registers.
219 // This can be tricky, because CFA register is ebp (unaligned)
220 // and we need to produce offsets relative to it.
221 "movaps %xmm0, (%esp)\n"
222 "movaps %xmm1, 16(%esp)\n"
223 "movaps %xmm2, 32(%esp)\n"
224 "movaps %xmm3, 48(%esp)\n"
226 "movl 4(%ebp), %eax\n" // Pass prev frame and return address
227 "movl %eax, 4(%esp)\n"
228 "movl %ebp, (%esp)\n"
229 "call " ASMPREFIX "X86CompilationCallback2\n"
231 "movaps 48(%esp), %xmm3\n"
232 CFI(".cfi_restore %xmm3\n")
233 "movaps 32(%esp), %xmm2\n"
234 CFI(".cfi_restore %xmm2\n")
235 "movaps 16(%esp), %xmm1\n"
236 CFI(".cfi_restore %xmm1\n")
237 "movaps (%esp), %xmm0\n"
238 CFI(".cfi_restore %xmm0\n")
239 "movl %ebp, %esp\n" // Restore ESP
240 CFI(".cfi_def_cfa_register esp\n")
242 CFI(".cfi_adjust_cfa_offset 12\n")
244 CFI(".cfi_adjust_cfa_offset -4\n")
245 CFI(".cfi_restore %ecx\n")
247 CFI(".cfi_adjust_cfa_offset -4\n")
248 CFI(".cfi_restore %edx\n")
250 CFI(".cfi_adjust_cfa_offset -4\n")
251 CFI(".cfi_restore %eax\n")
253 CFI(".cfi_adjust_cfa_offset -4\n")
254 CFI(".cfi_restore %ebp\n")
256 CFI(".cfi_endproc\n")
259 void X86CompilationCallback2(void);
261 _declspec(naked) void X86CompilationCallback(void) {
266 call X86CompilationCallback2
275 #else // Not an i386 host
276 void X86CompilationCallback() {
277 assert(0 && "Cannot call X86CompilationCallback() on a non-x86 arch!\n");
283 /// X86CompilationCallback - This is the target-specific function invoked by the
284 /// function stub when we did not know the real target of a call. This function
285 /// must locate the start of the stub or call site and pass it into the JIT
286 /// compiler function.
288 extern "C" void X86CompilationCallback2() {
289 assert(sizeof(size_t) == 4); // FIXME: handle Win64
290 intptr_t *RetAddrLoc = (intptr_t *)_AddressOfReturnAddress();
291 RetAddrLoc += 4; // skip over ret addr, edx, eax, ecx
292 intptr_t RetAddr = *RetAddrLoc;
294 extern "C" void X86CompilationCallback2(intptr_t *StackPtr, intptr_t RetAddr) {
295 intptr_t *RetAddrLoc = &StackPtr[1];
297 assert(*RetAddrLoc == RetAddr &&
298 "Could not find return address on the stack!");
300 // It's a stub if there is an interrupt marker after the call.
301 bool isStub = ((unsigned char*)RetAddr)[0] == 0xCD;
303 // The call instruction should have pushed the return value onto the stack...
305 RetAddr--; // Backtrack to the reference itself...
307 RetAddr -= 4; // Backtrack to the reference itself...
311 DOUT << "In callback! Addr=" << (void*)RetAddr
312 << " ESP=" << (void*)StackPtr
313 << ": Resolving call to function: "
314 << TheVM->getFunctionReferencedName((void*)RetAddr) << "\n";
317 // Sanity check to make sure this really is a call instruction.
319 assert(((unsigned char*)RetAddr)[-2] == 0x41 &&"Not a call instr!");
320 assert(((unsigned char*)RetAddr)[-1] == 0xFF &&"Not a call instr!");
322 assert(((unsigned char*)RetAddr)[-1] == 0xE8 &&"Not a call instr!");
325 intptr_t NewVal = (intptr_t)JITCompilerFunction((void*)RetAddr);
327 // Rewrite the call target... so that we don't end up here every time we
330 *(intptr_t *)(RetAddr - 0xa) = NewVal;
332 *(intptr_t *)RetAddr = (intptr_t)(NewVal-RetAddr-4);
336 // If this is a stub, rewrite the call into an unconditional branch
337 // instruction so that two return addresses are not pushed onto the stack
338 // when the requested function finally gets called. This also makes the
339 // 0xCD byte (interrupt) dead, so the marker doesn't effect anything.
341 ((unsigned char*)RetAddr)[0] = (2 | (4 << 3) | (3 << 6));
343 ((unsigned char*)RetAddr)[-1] = 0xE9;
347 // Change the return address to reexecute the call instruction...
355 TargetJITInfo::LazyResolverFn
356 X86JITInfo::getLazyResolverFunction(JITCompilerFn F) {
357 JITCompilerFunction = F;
359 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
360 !defined(_MSC_VER) && !defined(__x86_64__)
361 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
367 if (!X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1)) {
368 // FIXME: support for AMD family of processors.
369 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
370 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
371 if ((EDX >> 25) & 0x1)
372 return X86CompilationCallback_SSE;
377 return X86CompilationCallback;
380 void *X86JITInfo::emitFunctionStub(void *Fn, MachineCodeEmitter &MCE) {
381 // Note, we cast to intptr_t here to silence a -pedantic warning that
382 // complains about casting a function pointer to a normal pointer.
383 #if (defined(__i386__) || defined(i386) || defined(_M_IX86)) && \
384 !defined(_MSC_VER) && !defined(__x86_64__)
385 bool NotCC = (Fn != (void*)(intptr_t)X86CompilationCallback &&
386 Fn != (void*)(intptr_t)X86CompilationCallback_SSE);
388 bool NotCC = Fn != (void*)(intptr_t)X86CompilationCallback;
392 MCE.startFunctionStub(13, 4);
393 MCE.emitByte(0x49); // REX prefix
394 MCE.emitByte(0xB8+2); // movabsq r10
395 MCE.emitWordLE(((unsigned *)&Fn)[0]);
396 MCE.emitWordLE(((unsigned *)&Fn)[1]);
397 MCE.emitByte(0x41); // REX prefix
398 MCE.emitByte(0xFF); // jmpq *r10
399 MCE.emitByte(2 | (4 << 3) | (3 << 6));
401 MCE.startFunctionStub(5, 4);
403 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
405 return MCE.finishFunctionStub(0);
409 MCE.startFunctionStub(14, 4);
410 MCE.emitByte(0x49); // REX prefix
411 MCE.emitByte(0xB8+2); // movabsq r10
412 MCE.emitWordLE(((unsigned *)&Fn)[0]);
413 MCE.emitWordLE(((unsigned *)&Fn)[1]);
414 MCE.emitByte(0x41); // REX prefix
415 MCE.emitByte(0xFF); // callq *r10
416 MCE.emitByte(2 | (2 << 3) | (3 << 6));
418 MCE.startFunctionStub(6, 4);
419 MCE.emitByte(0xE8); // Call with 32 bit pc-rel destination...
421 MCE.emitWordLE((intptr_t)Fn-MCE.getCurrentPCValue()-4);
424 MCE.emitByte(0xCD); // Interrupt - Just a marker identifying the stub!
425 return MCE.finishFunctionStub(0);
428 /// relocate - Before the JIT can run a block of code that has been emitted,
429 /// it must rewrite the code to contain the actual addresses of any
430 /// referenced global symbols.
431 void X86JITInfo::relocate(void *Function, MachineRelocation *MR,
432 unsigned NumRelocs, unsigned char* GOTBase) {
433 for (unsigned i = 0; i != NumRelocs; ++i, ++MR) {
434 void *RelocPos = (char*)Function + MR->getMachineCodeOffset();
435 intptr_t ResultPtr = (intptr_t)MR->getResultPointer();
436 switch ((X86::RelocationType)MR->getRelocationType()) {
437 case X86::reloc_pcrel_word: {
438 // PC relative relocation, add the relocated value to the value already in
439 // memory, after we adjust it for where the PC is.
440 ResultPtr = ResultPtr -(intptr_t)RelocPos - 4 - MR->getConstantVal();
441 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
444 case X86::reloc_picrel_word: {
445 // PIC base relative relocation, add the relocated value to the value
446 // already in memory, after we adjust it for where the PIC base is.
447 ResultPtr = ResultPtr - ((intptr_t)Function + MR->getConstantVal());
448 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
451 case X86::reloc_absolute_word:
452 // Absolute relocation, just add the relocated value to the value already
454 *((unsigned*)RelocPos) += (unsigned)ResultPtr;
456 case X86::reloc_absolute_dword:
457 *((intptr_t*)RelocPos) += ResultPtr;